Services Approach Credentials Research Blog Start an engagement

CVE-2026-47754: Unauthenticated Path Traversal in Metacat 2.x

This is the companion to our write-up of CVE-2026-48114. Alongside that SQL injection, we reported a second critical flaw in Metacat 2.x through the Synack platform: an unauthenticated path traversal in the legacy read endpoint. One GET request from anyone on the network was enough to read any file the Tomcat process could read. We found it together with our colleagues at LVL 0x00, and it is now public as CVE-2026-47754 with a CVSS 3.1 score of 9.3.

Metacat is an open-source data repository platform from the National Center for Ecological Analysis and Synthesis (NCEAS) at UC Santa Barbara, used to store and share research data and its metadata. Like the SQL injection, this flaw lived in the legacy 1.x API that Metacat 3.0.0 removed, so current releases are not affected. This post covers what was wrong, how to tell whether you are exposed, and what to rotate if you were.

Identifier
CVE-2026-47754 (GHSA-m852-f287-7cgw)
Product
Metacat, a Java servlet application from NCEAS
Affected
2.x through 2.19.1, and all 1.x versions
Fixed in
3.0.0 and later
Entry point
The archiveEntryName parameter of action=read
Severity
9.3 Critical, CVSS:3.1/AV:N/AC:L/PR:N/UI:N/S:C/C:H/I:L/A:N
Weaknesses
CWE-22 (Path Traversal), CWE-862 (Missing Authorization)

If you run Metacat 2.x

Start here, then read on for the detail.

Upgrade. Metacat 3.0.0 removed the vulnerable code, and the advisory recommends moving to the current 3.x release (3.4.0 at the time it was published). Aim to stay current rather than just reach 3.x: the project's security policy supports only the latest release and does not backport fixes to older ones.

If you cannot upgrade today, the advisory describes a workaround: in Tomcat, remove or restrict the servlet-mapping entries in web.xml for edu.ucsb.nceas.metacat.MetaCatServlet, which are mapped to /metacat, /metacat/* and /servlet/metacat, then restart Tomcat. This switches off the whole 1.x API, so any client that still uses it will stop working. The advisory notes that the 3.x DataONE REST API provides equivalent capabilities. It also lists other 1.x servlets you can disable at the same time: ReplicationServlet, AdvancedSearchServlet, HarvesterServlet, DataProvider, HarvesterRegistrationLogin and HarvesterRegistration.

The CVE-2026-48114 workaround does not cover this one

The workaround for the SQL injection removes the harvester and OAI-PMH servlets. This flaw is in the main Metacat servlet, so if that was all you did, the file-read path is still open.

Check whether it was used. Unlike a POST body, the payload here travels in the query string, and Tomcat's standard access log records the full request line. Search your logs for requests to the Metacat servlet with action=read and an archiveEntryName value that contains .. or an encoded form such as %2e%2e, %2f or a double-encoded variant. Log retention and rotation limit how far back you can look, and a reverse proxy may log differently, so a clean result is reassuring but not proof.

If you cannot rule out use, treat everything the Tomcat user could read as exposed. Rotate the credentials in Metacat's configuration. Replace any client certificates and private keys the node uses to authenticate to the networks it belongs to, and tell the operators of those networks so the old certificates can be revoked. Consider whether unpublished or embargoed data on disk may have been disclosed.

As a stopgap, a reverse-proxy rule that rejects requests carrying an archiveEntryName parameter can help if none of your clients use it. Encodings make filter rules easy to bypass, so treat it as a delay and not a fix.

What was wrong

Two weaknesses combined, and neither is exotic. The legacy read handler trusted a path supplied by the client, and the permission check that should have run first was commented out.

1. A filesystem path built from user input

In action=read, the archiveEntryName parameter names an entry to return from a document's archive. ArchiveHandler.readArchiveEntry() joins that value onto a base directory without validating it. A value containing ../ segments climbs out of the base directory, and the handler then reads whatever it lands on. The pattern looks like this:

// Illustrative pattern only, not copied from the Metacat source
String path = archiveDir + "/" + archiveEntryName;  // "../" climbs out
return Files.readAllBytes(Paths.get(path));

The fix for this class of bug is to resolve the path, normalize it, and confirm that the result is still inside the base directory before opening anything:

Path base = archiveDir.toRealPath();
Path target = base.resolve(archiveEntryName).normalize();
if (!target.startsWith(base)) {
    throw new SecurityException("Entry is outside the archive directory");
}

Better still is never to accept a path from the client at all. Take an identifier and look the file up yourself.

2. The permission check was commented out

The code contains a hasReadPermission() check, but in this path it was commented out. That is CWE-862, and it is why the CVSS vector reads PR:N: no login and no privileges required. A single GET request from anyone who can reach the server is enough.

How the pieces fit together

1
Send one request
A GET to the read endpoint, with no login
2
Escape the folder
Dot-dot segments climb out of the archive directory
3
No check runs
The permission check is commented out
4
The file comes back
Anything the Tomcat user can read

Why it is rated critical

The severity comes from what the Tomcat process can read, which is more than the repository's documents. On a typical Metacat host that includes:

The S:C (scope changed) in the vector reflects that the exposure is not confined to Metacat. The files belong to the host and to the other systems that the stolen secrets unlock.

About the proof of concept

The proof of concept we shared with the maintainers read a standard, world-readable system file, which is enough to show that the traversal works. We are deliberately not publishing a ready-to-send request. The fix has been available since 3.0.0, but not every deployment has migrated, and the advisory gives defenders what they need to act.

The fix, and the gap before the advisory

Metacat 3.0.0, released in April 2024, removed the legacy API, including ArchiveHandler (commit 07e034b, PR #1713 for issue #1365). The commit and the issue describe architectural cleanup, not a security fix, so no advisory or CVE was issued. The 2.x branch was not backported and will not be.

The advisory was published in May 2026, more than two years after 3.0.0 shipped. In that time, nothing told operators of 2.x that their version let anyone read files off the server. The same was true of the SQL injection in the legacy 1.x servlets. Removing old code fixes the problem for people who upgrade. People who do not upgrade need to be told.

The maintainers took our reports seriously. They confirmed both issues within a week, drafted the advisories and invited us to collaborate on them, and contacted groups they knew were still on 2.x before the details became public. The project has also since merged a security policy that sends future reports through GitHub's private vulnerability reporting, so the next researcher has a documented route.

Disclosure timeline

  1. We submit the finding through the Synack platform to a customer program, together with LVL 0x00.
  2. Synack accepts the finding and the reward is paid.
  3. We report privately to the Metacat maintainers by email and propose 30 days before public disclosure.
  4. We verify the remediation on the affected customer deployment.
  5. The maintainers confirm the issue.
  6. We reply and offer to review draft advisories.
  7. The maintainers draft the advisory, invite us to collaborate on it, and propose a security policy for the repository.
  8. GitHub, acting as CNA, reserves CVE-2026-47754. The security policy is merged the same day.
  9. The maintainers publish GitHub advisory GHSA-m852-f287-7cgw.
  10. The CVE record and NVD entry are published. CISA's assessment at that point listed exploitation as "none", automatable as "yes" and technical impact as "partial".

The customer's deployment had been remediated, and the fix verified, before the identifier was reserved or the advisory published.

What to take from this

Credits

The GitHub advisory credits wwallace (Fjord.AI) and lvl0x00 as reporters. Thank you to the Metacat team, credited on the advisory as taojing2002 (remediation developer), iannesbitt (remediation verifier) and mbjones (coordinator), and to Synack and the customer's security team for their quick triage and remediation. Our write-up of the SQL injection we reported alongside this one is at CVE-2026-48114.

References

Find it before an adversary does

We find new vulnerabilities in the software you depend on, prove them with working proofs of concept, and coordinate disclosure with the vendor.

Start an engagement